Shared 1x1 Conv vs. Independent 1x1 Conv

[yyvettey]

Could you please explain the differences between these two in details? Would appreciated it if you could point out in the code. Many thanks!

[huanghoujing]

Hi, (Shared 1x1 Conv) means there is only one 1x1 Conv layer after pooling, shared between all parts/stripes (There are 6 parts, as an example). Independent 1x1 Conv means each part is processed by its 1x1 Conv layer, meaning different parts are processed by different model weights. The (Shared 1x1 Conv) implementation can only be found in early versions of code, not in current commit.

[abhishekaich27]

Hi @huanghoujing ! So now the implementation is for each stripe having it's own classifier as described in Figure 2 of the base paper?

[huanghoujing]

@abhishekaich27 Yes, you are right.